Skip to main content
Read this once before installing. If you can check every box in You are ready when…, the install will succeed.

​
Supported environment

“1.24+” is the recommended floor; it is not enforced by the chart. Very old clusters may work but are untested.

​
Resource footprint

Collector defaults (chart ≥ 5.3.6 / operator ≥ v5.3.5): requests 25m CPU / 64Mi memory, limits 500m CPU / 512Mi memory — grounded in measured usage on a small cluster (~9m CPU / 21Mi for 1 node, ~30 pods) with headroom, since usage scales with cluster size. These are the current defaults, validated on small clusters so far — recommended sizing will firm up as operational experience accumulates across medium, busy, and stress-scale fleets. Override any subset in standard Kubernetes shape, e.g. --set agent.collector.resources.limits.memory=1Gi. Sizing starting points: up to a few hundred pods, the defaults are comfortable; around a thousand pods, watch collector memory in your normal tooling; well beyond that, raise the memory limit to 1Gi and adjust from observation. Installs on older operator versions have no requests/limits set until upgraded.
No PersistentVolumeClaims are created. The collector runs non-root, and its RBAC is strictly read-only (get/list/watch) — see Permissions & RBAC.

​
Network

  • Outbound HTTPS (TCP 443) from the cluster to api.changeguard.ai (TLS 1.2+).
  • DNS resolution for api.changeguard.ai (and install.changeguard.ai if you use the one-line installer).
  • Image pulls from public.ecr.aws (or your mirror — the chart supports a custom global.imageRegistry for air-gapped setups).
No outbound HTTP proxy support today. The collector needs direct HTTPS egress; HTTP_PROXY / HTTPS_PROXY env vars are not honored by its push client. If your cluster only reaches the internet through a proxy, contact us before installing.

​
Permissions to install

You need rights to create namespaces, a CRD, ClusterRoles/Bindings, and Deployments. cluster-admin is not required, and ChangeGuard AI requests no wildcard RBAC. See Permissions for the exact grants (including that the operator and collector read Secrets and RBAC objects cluster-wide, read-only).

​
Credentials & secrets

  • A ChangeGuard AI API key (create one at app.changeguard.ai → API keys). Treat it like a password.
  • The chart creates one Kubernetes Secret, changeguard-credentials (key api-key), in changeguard-agent. You can also create it yourself and reference it.

​
Data collected vs. not collected

​
What the installer changes — and doesn’t

Creates: namespaces changeguard-system and changeguard-agent; the changeguardagents.changeguard.ai CRD; the operator (Deployment + RBAC); a ChangeGuardAgent custom resource; the collector Deployment; one Secret; optionally scanner CronJobs/DaemonSet. Does not change: any of your existing workloads, your RBAC for other principals, your Secrets, or your CI/CD — unless you later opt into remediation and apply its namespaced RBAC.

​
Supported platforms

Kubernetes 1.24+ on EKS, GKE, AKS, or any CNCF-conformant distribution, with Helm 3.8+ and kubectl. Analysis and scoring run in the control plane, so in-cluster requirements stay minimal.

​
Container images & registries

The cluster pulls images from these registries — allow them in egress policy, or mirror them for air-gapped installs: Air-gapped or mirrored-registry installs can repoint every image at one registry with --set global.imageRegistry=<your-registry>.

​
AI investigation & model access

The default install includes an operator-managed AI investigator (HolmesGPT) that calls your model provider. On EKS with Amazon Bedrock, grant it access via IRSA:
Without model access, AI investigation analyses won’t run — data collection, the CSC Score, and pre-flight checks are unaffected.

​
Upgrades & uninstall — plan for both before you start

Upgrades are standard Helm upgrades, and most improvements ship control-plane-side with no cluster change — see the Operations handbook. Uninstall removes the in-cluster footprint cleanly, with one manual step (deleting the CRD) — the exact commands are on the install page.

​
Known limitations — read before your security review

  • No HTTP(S) proxy support — the cluster needs direct outbound HTTPS to api.changeguard.ai:443.
  • kubectl get changeguardagent shows blank CSC/PUSHES columns — the dashboard is the source of truth for score and push activity.
  • Key rotation has no self-serve rotation API today; for high-assurance invalidation after a suspected exposure, rotate and confirm with support.
  • Data removal is per-cluster in-app; full-tenant erasure goes through support.
The consolidated, always-current list — including Early Access status per capability — is Known Limitations & Roadmap.

​
You are ready when…

  • kubectl points at the cluster you intend to install into (kubectl config current-context).
  • Helm 3.8+ is installed (helm version --short).
  • The cluster runs Kubernetes 1.24+.
  • You can create cluster-scoped resources (namespaces, a CRD, ClusterRoles) — cluster-admin is not required.
  • The cluster has outbound HTTPS to api.changeguard.ai:443 (no HTTP proxy).
  • You have a ChangeGuard AI API key (cg_…), created at app.changeguard.ai.
  • You’ve decided whether to keep security scanning on — it’s the default and includes a privileged Falco DaemonSet (see Permissions & RBAC).
  • (EKS + Bedrock) You know which IAM role the AI investigator will use for model access — see AI investigation & model access.
Next: Install →